[Paper Review] Investigating Bias in LLM-Based Bias Detection: Disparities between LLMs and Human Perception

Luyang Lin, Lingzhi Wang+2 more|arXiv (Cornell University)|Mar 22, 2024

Forecasting Techniques and Applications20 citations

TL;DR

The paper investigates inherent biases in large language models (LLMs) and how these biases affect media bias detection, examining political bias prediction and text continuation, topic consistency, debiasing strategies, and cross-model bias tendencies.

ABSTRACT

The pervasive spread of misinformation and disinformation in social media underscores the critical importance of detecting media bias. While robust Large Language Models (LLMs) have emerged as foundational tools for bias prediction, concerns about inherent biases within these models persist. In this work, we investigate the presence and nature of bias within LLMs and its consequential impact on media bias detection. Departing from conventional approaches that focus solely on bias detection in media content, we delve into biases within the LLM systems themselves. Through meticulous examination, we probe whether LLMs exhibit biases, particularly in political bias prediction and text continuation tasks. Additionally, we explore bias across diverse topics, aiming to uncover nuanced variations in bias expression within the LLM framework. Importantly, we propose debiasing strategies, including prompt engineering and model fine-tuning. Extensive analysis of bias tendencies across different LLMs sheds light on the broader landscape of bias propagation in language models. This study advances our understanding of LLM bias, offering critical insights into its implications for bias detection tasks and paving the way for more robust and equitable AI systems

Motivation & Objective

• Assess whether LLMs exhibit political bias in bias prediction and in text continuation.
• Examine bias consistency of LLMs across diverse topics (predefined and latent).
• Investigate debiasing strategies via prompt engineering and model fine-tuning and their effects on performance.
• Compare bias tendencies across multiple open-source and closed-source LLMs to understand model-variance in bias behavior.

Proposed method

• Use vanilla ChatGPT to evaluate political leaning on FlipBias and ABP datasets through a three-way label task (left, center, right, uncertain).
• Conduct article continuation experiments with prefixes from political articles and analyze generated suffixes using embedding-based similarity and left/right vocabulary matching.
• Introduce Bias Tendency Index (BTI-1, BTI-2) to quantify topic-level bias tendencies across datasets.
• Apply debiasing approaches including prompt-based explanations, few-shot prompts, and debiasing statements (DS), as well as fine-tuning with varying label distributions (L-FT, LC-FT, LCR-FT).
• Evaluate debiasing impact on overall bias prediction and topic-level bias distributions, reporting metrics like BiF1, MaF1, and BTI changes.
• Extend analysis to additional LLMs (LLaMa2, Vicuna, Mistral, GPT-4) to compare bias tendencies across models.

Experimental results

Research questions

• RQ1 RQ1: Do LLMs exhibit political bias in bias prediction and text continuation tasks?
• RQ2 RQ2: Do LLMs demonstrate consistent bias across topics (predefined and latent)?
• RQ3 RQ3: How to debias LLMs and further improve bias-detection performance?
• RQ4 RQ4: Do various LLMs exhibit similar bias tendencies across datasets and topics?

Key findings

• LLMs show a left-leaning cognitive bias in political bias prediction on FlipBias and ABP, with higher Left-Center proportions than Right-Center in the forecasts.
• LLMs perform better at predicting right-leaning ground-truth articles than left-leaning ones, suggesting asymmetric bias tendencies.
• Article continuation experiments indicate a left-leaning leaning for shorter prefixes, shifting toward right-leaning as prefix length increases due to topic length differences.
• Bias Tendency Index (BTI-1, BTI-2) reveals topic-dependent bias, with most topics showing left-leaning tendencies but notable right-leaning topics as well.
• Debiasing via prompt-based methods (especially Debiasing Statement) reduces bias alignment around zero for topic-level BTI, while fine-tuning can improve bias-prediction metrics but may increase overall bias.
• Different LLMs exhibit diverse bias profiles; some models show stronger or different bias directions, and model performance does not strictly correlate with bias strength.